The ggml‑org team published a pre‑release of llama.cpp labeled b11532. The update adds exact GELU activation for ModernBERT encoders and keeps tanh‑GELU aliases. It also maps the Python GELU function to a more precise implementation. The release includes builds for many platforms, from macOS Apple Silicon to Windows CUDA 13.4. The source code is signed with a verified GPG key.
Why it matters
Developers can run ModernBERT models locally with more accurate activation functions, improving inference quality without cloud services.
